Created by: heygrain

1）PaddlePaddle版本：1.6 2）GPU：GTX1070单卡

错误信息：

Traceback (most recent call last):
  File "/home/grain/.vscode/extensions/ms-python.python-2019.11.50794/pythonFiles/ptvsd_launcher.py", line 43, in <module>
    main(ptvsdArgs)
  File "/home/grain/.vscode/extensions/ms-python.python-2019.11.50794/pythonFiles/lib/python/old_ptvsd/ptvsd/__main__.py", line 432, in main
    run()
  File "/home/grain/.vscode/extensions/ms-python.python-2019.11.50794/pythonFiles/lib/python/old_ptvsd/ptvsd/__main__.py", line 316, in run_file
    runpy.run_path(target, run_name='__main__')
  File "/usr/lib/python3.6/runpy.py", line 263, in run_path
    pkg_name=pkg_name, script_name=fname)
  File "/usr/lib/python3.6/runpy.py", line 96, in _run_module_code
    mod_name, mod_spec, pkg_name, script_name)
  File "/usr/lib/python3.6/runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "/media/grain/DATA/Grain_JGY/WorkSpace/graph_cross/tmp/tmp.py", line 98, in <module>
    optimizer.minimize(model.loss)
  File "</home/grain/ml/lib/python3.6/site-packages/decorator.py:decorator-gen-36>", line 2, in minimize
  File "/home/grain/ml/lib/python3.6/site-packages/paddle/fluid/wrapped_decorator.py", line 25, in __impl__
    return wrapped_func(*args, **kwargs)
  File "/home/grain/ml/lib/python3.6/site-packages/paddle/fluid/dygraph/base.py", line 78, in __impl__
    return func(*args, **kwargs)
  File "/home/grain/ml/lib/python3.6/site-packages/paddle/fluid/optimizer.py", line 678, in minimize
    no_grad_set=no_grad_set)
  File "/home/grain/ml/lib/python3.6/site-packages/paddle/fluid/optimizer.py", line 551, in backward
    no_grad_set, callbacks)
  File "/home/grain/ml/lib/python3.6/site-packages/paddle/fluid/backward.py", line 1085, in append_backward
    _append_backward_vars_(root_block, fwd_op_num, grad_to_var, grad_info_map)
  File "/home/grain/ml/lib/python3.6/site-packages/paddle/fluid/backward.py", line 891, in _append_backward_vars_
    op_desc.infer_shape(block.desc)
paddle.fluid.core_avx.EnforceNotMet: 

--------------------------------------------
C++ Call Stacks (More useful to developers):
--------------------------------------------
0   std::string paddle::platform::GetTraceBackString<std::string const&>(std::string const&, char const*, int)
1   paddle::platform::EnforceNotMet::EnforceNotMet(std::string const&, char const*, int)
2   paddle::operators::Reshape2GradOp::InferShape(paddle::framework::InferShapeContext*) const
3   paddle::framework::OpDesc::InferShape(paddle::framework::BlockDesc const&) const

------------------------------------------
Python Call Stacks (More useful to users):
------------------------------------------
  File "/home/grain/ml/lib/python3.6/site-packages/paddle/fluid/framework.py", line 2488, in append_op
    attrs=kwargs.get("attrs", None))
  File "/home/grain/ml/lib/python3.6/site-packages/paddle/fluid/layer_helper.py", line 43, in append_op
    return self.main_program.current_block().append_op(*args, **kwargs)
  File "/home/grain/ml/lib/python3.6/site-packages/paddle/fluid/layers/nn.py", line 9022, in reshape
    "XShape": x_shape})
  File "/media/grain/DATA/Grain_JGY/WorkSpace/graph_cross/tmp/tmp.py", line 72, in __init__
    concat_users = layers.reshape(layers.concat([self.users] * (n_negsamples + 1)), [-1, 1], inplace=False)
  File "/media/grain/DATA/Grain_JGY/WorkSpace/graph_cross/tmp/tmp.py", line 88, in <module>
    model = GraphCross(user_id_range=data['user_id_range'], item_id_range=data['item_id_range'], n_negsamples=3)
  File "/usr/lib/python3.6/runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "/usr/lib/python3.6/runpy.py", line 96, in _run_module_code
    mod_name, mod_spec, pkg_name, script_name)
  File "/usr/lib/python3.6/runpy.py", line 263, in run_path
    pkg_name=pkg_name, script_name=fname)
  File "/home/grain/.vscode/extensions/ms-python.python-2019.11.50794/pythonFiles/lib/python/old_ptvsd/ptvsd/__main__.py", line 316, in run_file
    runpy.run_path(target, run_name='__main__')
  File "/home/grain/.vscode/extensions/ms-python.python-2019.11.50794/pythonFiles/lib/python/old_ptvsd/ptvsd/__main__.py", line 432, in main
    run()
  File "/home/grain/.vscode/extensions/ms-python.python-2019.11.50794/pythonFiles/ptvsd_launcher.py", line 43, in <module>
    main(ptvsdArgs)

----------------------
Error Message Summary:
----------------------
Error: Input(Out@GRAD) shouldn't be null.
  [Hint: Expected ctx->HasInput(framework::GradVarName("Out")) == true, but received ctx->HasInput(framework::GradVarName("Out")):0 != true:1.] at (/paddle/paddle/fluid/operators/reshape_op.cc:470)
  [operator < reshape2_grad > error]

可供调试的代码：

import numpy as np
import paddle
import paddle.fluid as fluid
import paddle.fluid.layers as layers
from scipy.sparse import coo_matrix

IS_SPARSE = True

# ===========这个函数是创建并加载一个调试用的模拟数据集==========
def load_toy_data():
    item_demo = np.array([8, 8, 8, 8]).reshape((-1, 1))

    train_user = np.array([0, 1, 2, 3, 0, 1, 2, 3]).reshape((-1, 1))
    train_item = np.array([4, 5, 6, 7, 5, 4, 7, 6]).reshape((-1, 1))

    data = dict()
    data['train_user'] = train_user
    data['train_item'] = train_item
    data['item_demo'] = item_demo
    data['user_demo'] = None
    data['item_idx_min'] = 4
    data['item_idx_max'] = 7
    data['user_id_range'] = [0, 3]
    data['item_id_range'] = [4, 7]

    n_fields = 3
    n_features = 9

    # user-item graph (only use training interactions)
    adj = np.hstack([data['train_user'], data['train_item']])
    # user-demo graph
    if data['user_demo'] is not None:
        adj = np.vstack([adj, np.hstack([np.arange(0, data['item_idx_min']).reshape((-1, 1)), data['item_demo']])])
    # item-demo graph
    if data['item_demo'] is not None:
        adj = np.vstack([adj, np.hstack([np.arange(data['item_idx_min'], data['item_idx_max'] + 1).reshape((-1, 1)), data['item_demo']])])

    adj_mat = coo_matrix((np.ones(len(adj)), (adj[:, 0], adj[:, 1])), shape=(n_features, n_features))
    adj_mat += adj_mat.transpose()
    data['adj_mat'] = adj_mat

    print('toy dataset:')

    for key, val in data.items():
        if (val is not None) and not (key in ['item_idx_min', 'item_idx_max', 'user_id_range', 'item_id_range']):
            print('\t', key, ':', val.shape)
    print('\tn_fields:', n_fields)
    print('\tn_nodes/feats:', n_features)

    return data, n_fields, n_features


class GraphCross(object):

    def __init__(self, user_id_range, item_id_range, n_negsamples):
        # ===============================model config===============================
        self.user_id_range = user_id_range
        self.item_id_range = item_id_range
        self.n_users = user_id_range[1] - user_id_range[0] + 1
        self.n_items = item_id_range[1] - item_id_range[0] + 1

        # =================================inputs===============================
        self.users = fluid.data(name='users', shape=[-1, 1], dtype='int64')
        self.pos_items = fluid.data(name='pos_items', shape=[-1, 1], dtype='int64')
        self.neg_items = fluid.data(name='neg_items', shape=[-1, n_negsamples], dtype='int64')
        batch_size = layers.shape(self.users)[0]
        self.batch_size = batch_size

        # ===============================negative samples transfer===============================
        concat_items = layers.reshape(layers.concat([self.pos_items, self.neg_items], axis=1), [-1, 1], inplace=False)
        concat_items = layers.elementwise_sub(concat_items, layers.fill_constant(shape=layers.shape(concat_items), dtype='int64', value=self.item_id_range[0]))
        concat_users = layers.reshape(layers.concat([self.users] * (n_negsamples + 1)), [-1, 1], inplace=False)

        u_vec = fluid.embedding(concat_users, [self.n_users, 5], is_sparse=IS_SPARSE)  # (N, embedding_dim=5)
        i_vec = fluid.embedding(concat_items, [self.n_items, 5], is_sparse=IS_SPARSE)
        scores = layers.reduce_sum(layers.elementwise_mul(u_vec, i_vec), dim=1)  # of shape (N, )
        scores = layers.reshape(scores, [batch_size, -1], inplace=False)  # of shape (batch_size, (n_negsamples + 1))

        # ===============================loss and metrics===============================
        bpr = layers.bpr_loss(input=scores, label=layers.fill_constant(shape=[batch_size, 1], value=0, dtype='int64'))
        self.loss = layers.mean(bpr)


# load data
data, n_fields, n_features = load_toy_data()

# create model
model = GraphCross(user_id_range=data['user_id_range'], item_id_range=data['item_id_range'], n_negsamples=3)

# test program
test_program = fluid.default_main_program().clone(for_test=True)

# creat optimizer
optimizer = fluid.optimizer.Adam(
    learning_rate=0.001,
    regularization=fluid.regularizer.L2DecayRegularizer(1e-5)
)
optimizer.minimize(model.loss)

# use cuda or not, and startup
use_cuda = True
place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())

train_user, train_item = data['train_user'], data['train_item']
neg_item = np.random.randint(low=data['item_idx_min'], high=data['item_idx_max'] + 1, size=[len(train_user), 3])  # [n_users, n_negsamples]
output = exe.run(feed={model.users.name: train_user.astype(np.int64),
                       model.pos_items.name: train_item.astype(np.int64),
                       model.neg_items.name: neg_item.astype(np.int64)}, 
                 fetch_list=[model.loss])

问题所在：

优化器那行，即： optimizer.minimize(model.loss) 注释掉之后，可以跑通得到model.loss， 但是： 有optimizer，产生反向传播就不行。 应该是对输入self.users、self.pos_item、self.neg_item进行concat、reshape等操作时导致了梯度计算的异常，请问应该怎么处理？

换句话说： 在自然语言处理(或者推荐系统)等场景，假设我的输入是词的id(或者用户和商品id)，这时候我要先对fluid.data接受的id矩阵（dtype='int64'）做一些concat和reshape的操作，然后再拿预处理完的id矩阵去做嵌入，也就是fluid.embedding。 反向传播梯度的问题就出在一开始预处理id的concat和reshape上面，实际上我并不需要梯度反向传播到这里，因为相当于预处理完了的数据才是真正的输入。所以，有适合用的api吗，不然我只能在cpu里用numpy先concat和reshape完了再feed模型了。。。